x_mu = tf.placeholder(shape=[None, 2], dtype=tf.float32)
x_theta = tf.placeholder(shape=[None, 2, N_GEN], dtype=tf.float32)
inds_gen = tf.placeholder(
shape=[None, 2, N_GEN],
dtype=tf.float32) # determines which generator is chosen for each
# of the samples.
# should be constant along the second dimension,
# with entries either 0 or 1 so that
# inds[i, j, :] sums to one
y = 0
for i in range(N_GEN):
y += inds_gen[:, :, i] * ff_net(x_theta[:, :, i],
'T_' + str(i),
input_dim=2,
output_dim=2,
activation=ACT_T,
n_layers=LAYERS_T,
hidden_dim=HIDDEN_T)
T_theta = y
h_mu = 0 # sum over h evaluated at samples of mu
h_T_theta = 0 # sum over h evaluated at samples of theta
for i in range(2):
h_mu += ff_net(x_mu[:, i:(i + 1)],
'h' + str(i),
input_dim=1,
output_dim=1,
activation=ACT_H,
n_layers=LAYERS_H,
hidden_dim=HIDDEN_H)
return -tf.reduce_sum(tf.square(u[:, 0:D_EACH] - u[:, D_EACH:D]), axis=1)
# build tensorflow graph
for run_K in range(N_RUNS):
t0 = time.time()
tf.reset_default_graph()
#
x_mu = tf.placeholder(shape=[None, D], dtype=tf.float32) # samples from mu
x_theta = tf.placeholder(shape=[None, D_THETA],
dtype=tf.float32) # samples from theta
#
T_theta = ff_net(x_theta,
'T',
input_dim=D_THETA,
output_dim=D,
activation=ACT_T,
n_layers=LAYERS_T,
hidden_dim=HIDDEN_T)
#
h_mu = 0 # sum over h evaluated at samples of mu
h_T_theta = 0 # sum over h evaluated at samples of theta
for i in range(2):
h_mu_h = ff_net(x_mu[:, i * D_EACH:(i + 1) * D_EACH],
'h_' + str(i),
input_dim=D_EACH,
output_dim=1,
activation=ACT_H,
n_layers=LAYERS_H,
hidden_dim=HIDDEN_H)
h_mu += h_mu_h # + psi_star(h_mu_h)
# build tensorflow graph
for run_K in range(N_RUNS):
t0 = time.time()
tf.reset_default_graph()
x_mu = tf.placeholder(shape=[None, 2], dtype=tf.float32) # samples from mu
x_theta = tf.placeholder(shape=[None, 2],
dtype=tf.float32) # samples from theta
x_kappa = tf.placeholder(shape=[None, 1],
dtype=tf.float32) # samples from kappa
T_theta = ff_net(x_theta,
'T',
input_dim=2,
output_dim=2,
activation=ACT_T,
n_layers=LAYERS_T,
hidden_dim=HIDDEN_T)
h_mu = 0 # sum over h evaluated at samples of mu
h_T_theta = 0 # sum over h evaluated at samples of theta
for i in range(2):
h_mu += ff_net(x_mu[:, i:(i + 1)],
'h_' + str(i),
input_dim=1,
output_dim=1,
activation=ACT_H,
n_layers=LAYERS_H,
hidden_dim=HIDDEN_H)
h_T_theta += ff_net(T_theta[:, i:(i + 1)],
def f_objective(u):
return tf.nn.relu(tf.reduce_sum(u, axis=1))
# build tensorflow graph
for run_K in range(N_RUNS):
t0 = time.time()
tf.reset_default_graph()
x_mu = tf.placeholder(shape=[None, 2], dtype=tf.float32) # samples from mu
x_theta = tf.placeholder(shape=[None, 2], dtype=tf.float32) # samples from theta
x_kappa = tf.placeholder(shape=[None, 1], dtype=tf.float32) # samples from kappa
T_theta = ff_net(x_theta, 'T', input_dim=2, output_dim=2, activation=ACT_T, n_layers=LAYERS_T, hidden_dim=HIDDEN_T)
h_mu = 0 # sum over h evaluated at samples of mu
h_T_theta = 0 # sum over h evaluated at samples of theta
for i in range(2):
h_mu_h = ff_net(x_mu[:, i:(i + 1)], 'h_' + str(i), input_dim=1, output_dim=1, activation=ACT_H,
n_layers=LAYERS_H, hidden_dim=HIDDEN_H)
h_mu += h_mu_h + psi_star(h_mu_h)
h_T_theta += ff_net(T_theta[:, i:(i + 1)], 'h_' + str(i), input_dim=1, output_dim=1, activation=ACT_H,
n_layers=LAYERS_H, hidden_dim=HIDDEN_H)
h_kappa_mu_h = ff_net(x_kappa, 'h_kappa', input_dim=1, output_dim=1, activation=ACT_H, n_layers=LAYERS_H,
hidden_dim=HIDDEN_H)
h_kappa_mu = h_kappa_mu_h + psi_star(h_kappa_mu_h)
h_kappa_T_theta = ff_net(T_theta[:, 0:1] - T_theta[:, 1:2], 'h_kappa', input_dim=1, output_dim=1,
activation=ACT_H, n_layers=LAYERS_H, hidden_dim=HIDDEN_H)